Dynamic self-configurable architectures for intelligent learning agents

GOAL

The research team, led by Prof. Ron Meir (Technion) and Prof. Naftali Tishby (HUJI),consists of researchers from the Technion and Hebrew University with broad expertise in neuroscience, neural computation and machine learning, VLSI systems, computer architectures, distributed systems, and networks.

The overall goal of this project is to contribute to the development of energy-efficient self-configurable adaptive and distributed learning architectures which are able to perceive and act in the world in real time. The mean for that is to design highly parallel robust-to-errors processors that modify their internal structure to adapt themselves to their working environment. This new type of complex and robust computing systems will be built from simple and fragile (e.g., small size, low cost, low power) building blocks. These artificial systems will continually sample their environment (e.g., video, audio), modify their internal structure at multiple spatial and temporal scales, point to regions of interest in large scale data streams, and achieve well-defined functional goals. Similarly to biological networks, the envisaged computational structure will be based on a self-configurable hierarchical architecture, where the computation and memory co-reside in the same elements, and where the hardware itself adapts to the task at hand. For this goal, we plan doing research in the following areas:

• Neuromorphic systems for Pattern recognition and control

• Brain-inspired devices and memristors –based circuits

• Neural-inspired network architecture

This team will focus on integrating neuromorphic hardware and brain inspired principles and algorithms. Applications that can benefit from this research include audio/visual systems, natural language processing, and perception-action integration, learning and control.

Brain-inspired devices and circuit construction (Avinoam Kolodny)

The group explores memristor devices as potential basic elements for building brain-inspired circuits and systems. Memristors are passive two-terminal devices which can be fabricated at high density within integrated CMOS circuits. They can be designed and configured to function as memory elements in array structures, as nonlinear components in logic gates, or as analog integrators. Current ongoing research is focused on memristor-based circuits for implementing Hebbian learning (Soudry, Kvatinsky, Di-Castro, Gal, Kolodny), and on programmable logical cell arrays (Kvatinsky, Yifat Levy, Cassuto, Kolodny in collaboration with Eitan Yaakobi and Shuki Bruck in Caltech).

Current students: Shahar Kvatinsky, Daniel Soudry, Yifat Levy

Publications

– Journal paper In preparation: “Hebbian Learning Rules with Memristors” – Daniel Soudry, Shahar Kvatinsky, Asaf Gal, Dotan Di Castro and Avinoam Kolodny

Ex-vivo developing cortical network growing on a micro-electrode array. The networks can be stimulated and activities of neurons can be measured at 60-120 positions. Behaviors of both single neurons and networks can be studied, and different networks can be combined in a modular fashion.

Neuroscience (Shimon Marom, Noam Ziv and Eilon Vaadia)

The work in Marom’s group is directed towards (1) identifying single neuron and neural network dynamical and functional characteristics, in the biological context; and (2) casting these characteristics in terms translatable to neural-inspired architectures and devices.

Ziv’s lab has been studying spontaneous, activity dependent and independent synaptic remodeling to explore the hypothesis that spontaneous synaptic changes reflect “exploratory” process that allow networks to sample synaptic configuration space  from which  functionally “useful” configurations are selected. To that end the lab combines recordings from networks of cortical neurons growing on multielectrode arrays and automatic microscopy of synaptic remodeling to study relationships between activity and synaptic remodeling and use and closed loop controllers to examine the capacity of external cues to stabilize selected synaptic configurations according to predefined criteria.

Current students:  Asaf Gal (postdoc), Maya Kaufman, Roman Dvorkin, Anna Rubinski

Publications

  1. R. Vardi, R. Timor, S. Marom, M. Abeles, and I. Kanter. “Synchronization by elastic neuronal latencies”. Phys. Rev. E 87, 012724A, doi: 10.1103/PhysRevE.87.012724, 2013.
  2. A. Gal and S. Marom, “(2013) Entrainment of the intrinsic dynamics of single isolated neurons by natural-like input”. The Journal of Neuroscience 33(18):7912-7918
  3. A. Gal and S. Marom, “(2013) Self-organized criticality in single neuron excitability”. arXiv:1210.7414 [q-bio.NC
  4. A. Gal and S. Marom, “Single Neuron Response Fluctuations: a Self-Organized Criticality Point of View. In: “Criticality in Neural Systems”, edited by Ernst Niebur, Dietmar Plenz, and Heinz G. Schuster, Wiley Blackwell (Berlin), in-press
  5. Cohen LD, Zuchman R, Sorokina O, Müller A, Dieterich DC, Armstrong JD, Ziv T, Ziv NE. “(2013) Metabolic turnover of synaptic proteins: kinetics, interdependencies and implications for synaptic maintenance”. PLoS One. 8(5):e63191
  6. Fisher-Lavie A, Ziv NE, “(2013) Matching dynamics of pre and postsynaptic scaffolds” (under review)
  7. Asaf Gal and Shimon Marom, “Self-organized criticality in single neuron excitability”.  Physical Review E, 88 (2013)
  8. Asaf Gal and Shimon Marom, “Entrainment of the intrinsic dynamics of single isolated neurons by natural-like input”.  The Journal of Neuroscience 33(18):7912-7918 (2013).
  9. Roni Vardi, Reut Timor, Shimon Marom, Moshe Abeles, and Ido Kanter, “Synchronization by elastic neuronal latencies”. Physical Review E, 87, 012724 (2013).

Pattern recognition and control in neuromorphic networks (Ron Meir, Naftali Tishby and Amir Globerson)

Meir’s group studies biologically inspired agents, sensing and acting within an environment. The focus is on the representation of environmental signals within distributed networks of fault-prone elements, and on the interaction between sensory and motor activities. More specifically, they consider unsupervised neural self-organization within a sensorimotor loop, as well as reinforcement learning paradigms leading to goal-directed behavior. The aim of the research is to understand how distributed networks of biologically inspired entities self-organize themselves in order to solve well defined computation goals. The work will rely on, and interact with, the ongoing closed-loop experimental research taking place in the Marom and Ziv labs.

Current students:  Daniel Soudry

Publications

  1. D. Soudry and R. Meir, “The neuron’s response at extended timescales: a linearized spiking input-output relation”Front. Comput. Neurosci., 8(29), 2014
  2. D. Soudry and R. Meir, “The neuron’s response at extended timescales:long-term correlations without long-term memory” , Front. Comput. Neurosci., 8(35), 2014
  3. D. Soudry and R. Meir, “Spiking input-output relation for general biophysical neuron models”, Arxiv preprint arXiv:1207.4436, 2012 – arxiv.org

Neurally inspired networks (Isaac Keslassy, Idit Keidar, Ariel Orda)

The robust-networks group is concerned with mimicking the salient properties of neural networks to inspire fault-tolerant computing and networking architectures. The robust-networks group is focusing on two directions: (a) robust computation, and (b) robust routing. In particular, they consider how a network can react differently and adapt to different inputs in order to provide a stronger fault-tolerance and reliability. They are particularly interested in distributed approaches, where each node of the network can adopt a local policy while still achieving efficient overall network reliability.

Current students:  Asa Dan, Shay Vargaftik

PEOPLE
Prof. Ron Meir, Technion EE
Prof. Naftali Tishby, HUJI CSE
Dr. Daniel Ben-Dayan Rubin, Intel
Dr. Boris Ginzburg, Intel
Prof. Amir Globerson, HUJI CSE
Prof. Omri Barak, Technion Medicine
Prof. Isaac Keslassy, Technion EE
Prof. Avinoam Kolodny, Technion EE
Prof. Shimon Marom, Technion Medicine and EE
Prof. Ariel Orda, Technion EE
Prof. Eilon Vaadia, HUJI
Prof. Noam Ziv, Technion Medicine